-
- GC43G-0869: Evaluating a Deep Learning Weather Prediction Model with a Baroclinic Wave Test on an Aquaplanet
-
Board 0869‚ Hall EFG (Poster Hall)NOLA CC
Author(s):Generic 'disconnected' Message
William Yik, University of Washington Seattle Campus (First Author, Presenting Author)
Dale Durran, University of Washington Seattle Campus
Deep learning weather prediction models have shown great success in recent years, rivaling our best traditional weather models. However, given the black box nature of deep learning, it is not yet clear if these models have actually learned fundamental atmospheric physics. To test this, we apply one particular deep learning weather prediction model known as DLESyM to an aquaplanet, a hypothetical planet with no continents. Since the model has only been trained on historical weather data, the aquaplanet is an extremely challenging out-of-distribution test case which can reveal what the model has learned about atmospheric dynamics. We find that DLESyM is able to make reasonable predictions in this aquaplanet environment, with its results qualitatively similar to a traditional weather model. Other deep learning weather prediction models, however, quickly revert back to the real world weather conditions they were trained on, despite being initialized on an aquaplanet. We hypothesize that DLESyM’s success compared to other deep learning models is attributable to specific design choices in its neural network architecture.
Scientific DisciplineSuggested ItinerariesNeighborhoodType
Enter Note
Go to previous page in this tab
Session
